Research Deep-Dive

Planner vs Reactive Agent Architectures

信頼度 中比較編集: Cairni · 방금 · AI 生成v1

Overview

Two broad architectural camps have emerged for building LLM agents: explicit planners and reactive loops. Both extend a base LLM beyond a single prompt, but they differ fundamentally in how they sequence reasoning and action. LLM Agents — Research Overview frames this as one of the central open questions in the field. Research — LLM Agents.md


Explicit Planners

In a planner architecture, the agent first decomposes a goal into a sequence of sub-tasks, then executes them in order. The plan is produced upfront, before any tools are called or observations are made.

Strengths:

  • More predictable and auditable — you can inspect the plan before execution begins.
  • Easier to reason about correctness for well-defined, stable tasks.

Weaknesses:

  • Brittle when the world changes mid-task. If an early step produces unexpected output, the rest of the plan may be invalidated and the agent may not recover gracefully. Research — LLM Agents.md

Reactive Loops (ReAct-style)

Reactive agents do not plan ahead. Instead, they interleave reasoning and action at each step — think, act, observe, repeat. This is the core pattern described in ReAct Pattern.

Strengths:

  • Adapts dynamically to new information obtained from tool calls or environment observations.
  • Better suited to tasks where the path to the goal is not known in advance.

Weaknesses:

  • Can wander — without a structured goal decomposition, the agent may lose track of the original objective or get stuck in repetitive loops. Research — LLM Agents.md
  • The looping failure mode is also documented on ReAct Pattern: the model can repeat a failing action rather than backtrack.

Side-by-Side Comparison

DimensionExplicit PlannerReactive Loop
StructureDecompose, then executeInterleave reasoning and action
PredictabilityHighLow
AdaptabilityLow (brittle to change)High
Failure modePlan invalidated mid-taskWandering or looping
Best forStable, well-defined tasksDynamic, exploratory tasks

Research — LLM Agents.md


Architectural Diagram


Relationship to Other Concepts

  • ReAct Pattern is the canonical reactive architecture — see that page for a deeper look at the reason-act-observe cycle and its known failure modes.
  • Tool Use & Function Calling is central to both architectures; planners and reactive loops alike depend on reliable tool selection, which has its own challenges.
  • Agent Memory interacts with both: planners may store the plan itself, while reactive agents may rely more heavily on retrieval to maintain context across steps.
  • Chain-of-Thought Reasoning underpins both approaches but is extended in different ways — planners use it for upfront decomposition, reactive loops use it at each step.

Open Question

When is each architecture actually worth it?

This is explicitly flagged as unresolved in the source notes. See Open Questions for the full list of unresolved issues. Research — LLM Agents.md